Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

The Power of Privilege: Enhancing Land Cover Classification with Privileged Information

Eriksson, Agnes LU and Åhman, Malte (2024) In Master’s Theses in Mathematical Sciences FMAM05 20241
Mathematics (Faculty of Engineering)
Abstract
Given the pressing environmental crisis, marked by climate change, biodiversity loss, and deforestation, precise land cover monitoring has become crucial. In this thesis, we explore the use of privileged information to enhance land cover classification using deep neural networks. Our study specifically focuses on how to best utilise all available data, by leveraging a machine learning framework known as learning using privileged information, or LUPI for short. The main idea of the LUPI paradigm is to exploit extra information during training, which might not be available during inference. Based on the the recently released FLAIR 2 dataset, we evaluate three main LUPI strategies, of which a teacher student setup yields the most promising... (More)
Given the pressing environmental crisis, marked by climate change, biodiversity loss, and deforestation, precise land cover monitoring has become crucial. In this thesis, we explore the use of privileged information to enhance land cover classification using deep neural networks. Our study specifically focuses on how to best utilise all available data, by leveraging a machine learning framework known as learning using privileged information, or LUPI for short. The main idea of the LUPI paradigm is to exploit extra information during training, which might not be available during inference. Based on the the recently released FLAIR 2 dataset, we evaluate three main LUPI strategies, of which a teacher student setup yields the most promising results. Our proposed approach achieves an mIoU of 0.606, compared to the baseline score 0.588. These scores are also on par with the relevant state-of-the-art models. The teacher student architecture is mainly based on a U-Net with EfficientNet B4 as an encoder, but the results also generalise to other networks. In conclusion, we successfully exploit the available privileged information during training to improve inference results. (Less)
Please use this url to cite or link to this publication:
author
Eriksson, Agnes LU and Åhman, Malte
supervisor
organization
course
FMAM05 20241
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Machine Learning, Neural Networks, LUPI, Learning Using Privileged Information, Land Cover Classification, Semantic Segmentation, AI
publication/series
Master’s Theses in Mathematical Sciences
report number
LUTFMA-3535-2024
ISSN
1404-6342
other publication id
2024:E23
language
English
id
9166511
date added to LUP
2024-09-03 09:01:02
date last changed
2024-09-03 09:01:02
@misc{9166511,
  abstract     = {{Given the pressing environmental crisis, marked by climate change, biodiversity loss, and deforestation, precise land cover monitoring has become crucial. In this thesis, we explore the use of privileged information to enhance land cover classification using deep neural networks. Our study specifically focuses on how to best utilise all available data, by leveraging a machine learning framework known as learning using privileged information, or LUPI for short. The main idea of the LUPI paradigm is to exploit extra information during training, which might not be available during inference. Based on the the recently released FLAIR 2 dataset, we evaluate three main LUPI strategies, of which a teacher student setup yields the most promising results. Our proposed approach achieves an mIoU of 0.606, compared to the baseline score 0.588. These scores are also on par with the relevant state-of-the-art models. The teacher student architecture is mainly based on a U-Net with EfficientNet B4 as an encoder, but the results also generalise to other networks. In conclusion, we successfully exploit the available privileged information during training to improve inference results.}},
  author       = {{Eriksson, Agnes and Åhman, Malte}},
  issn         = {{1404-6342}},
  language     = {{eng}},
  note         = {{Student Paper}},
  series       = {{Master’s Theses in Mathematical Sciences}},
  title        = {{The Power of Privilege: Enhancing Land Cover Classification with Privileged Information}},
  year         = {{2024}},
}